Abstract: Considering the current timeconsuming feature extraction of the braincomputer interface， a feature extraction method based on constrained ICA was proposed for P300BCI. The temporal P300 character of every subject was studied using the EEG image， and then， reference signals were built according to the temporal P300 character. Using the reference signals combined with ICA， the most correlative independent components were extracted based on 64channel EEG. According to the extracted independent components， 3dimensional feature vectors were built and put into the linear classifier at last. Two public datasets of BCI Competition II and III were used to verify the method. The results show that the recognition accuracy can be improved to 671% only with three times average， and to 952% with fifteen times average. The computation time is also shorter than other methods in the same experimental conditions.

Key words: braincomputer interface, electroencephalogram(EEG), feature extraction, constrained ICA, recognition accuracy